Strategic Reasoning on HardCore Don'tSayIt OOD (held-out variant)
22.92Win RateDEPT
Evaluation Results
| Method | Links | |
|---|---|---|
| DEPTBackbone=Qwen3-4B-Base2026.05 | 22.92 | |
| GRPOBackbone=Qwen3-4B-Base2026.05 | 22.01 | |
| DEPTBackbone=Qwen3-8B-Base2026.05 | 19.27 | |
| SPIRALBackbone=Qwen3-8B-Base2026.05 | 18.88 | |
| SPIRALBackbone=Qwen3-4B-Base2026.05 | 17.97 | |
| SPAGBackbone=Qwen3-8B-Base2026.05 | 12.12 | |
| MARSBackbone=Qwen3-4B-Base2026.05 | 12.11 | |
| GRPOBackbone=Qwen3-8B-Base2026.05 | 10.03 | |
| MARSBackbone=Qwen3-8B-Base2026.05 | 8.33 | |
| SPAGBackbone=Qwen3-4B-Base2026.05 | 7.03 | |
| VANILLABackbone=Qwen3-8B-Base2026.05 | 2.34 | |
| VANILLABackbone=Qwen3-4B-Base2026.05 | 0.39 |